feat(sp21): T2.2 Phase 8.6 — curriculum weights via z-score exp(-sharpe) (atomic)

v8 smoke (train-96wfk, commit 5694eb4df) cycle 1 confirmed
curric_conc=0.0000 despite Phase 8.2 (relaxed clamp) and 8.4 (CV
formula) — root cause was the per-segment weight function itself,
not the post-hoc concentration metric.

Old:  (1.0 / sharpe.max(0.01)).clamp(0.01, 100.0)
      → saturates at 100 when sharpe < 0.01
      → with PF<1 policy, every segment has |sharpe| < 0.01
      → all weights = 100 → uniform → CV=0

New:  exp(-(sharpe - sharpe_mean) / sharpe_std).clamp(-3, 3)
      → z-score normalised, scale-invariant
      → monotonic, smooth, no saturation cliff
      → preserves differential difficulty at any scale
      → uniform-input → std=0 → z=0 → uniform weights (cold-start ✓)

The Phase 8.4 CV formula was correct but operates on post-saturated
weights which destroy differential signal upstream. 8.6 fixes the
producer, 8.4 + 8.6 compose to give a meaningful curric_conc on
volume bars.

Note: win_conc=0 is intentional under PF<1 policy (compute_winner_concentration
guards against undefined "winner concentration" when policy is losing).
This is honest reporting, not a bug; addressed in audit doc rather
than this commit.

Pearls honoured:
  - pearl_controller_anchors_isv_driven: normalisation anchor =
    observed sharpe_std (signal-driven), not magic
  - feedback_isv_for_adaptive_bounds: z-clamp [-3, 3] is 3σ tail-
    cutoff for numerical safety, structural not tuned
  - pearl_first_observation_bootstrap: uniform Sharpe → uniform
    weights (cold-start sentinel preserved)
  - feedback_no_quickfixes: structural reformulation (1/x → exp(-z))
    not a clamp adjustment

Verification:
  cargo check -p ml --features cuda  # clean

Expected v9 cycle 1 curric_conc: ~0.3-0.6 (strong differential signal)
from typical per-segment Sharpe spread ~1σ across 8 segments.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/trainer/enrichment.rs

@@ -586,30 +586,55 @@ fn compute_hindsight_labels(trades: &[EvalTrade], pnl_std: f32) -> Vec<Hindsight
         .collect()
 }
 
-/// E8: Curriculum weights — inverse-Sharpe per segment.
+/// E8: Curriculum weights — z-score-normalised exp(-sharpe) per segment.
 ///
-/// **Phase 8.2 (2026-05-11)** — relaxed inverse-Sharpe clamp from
-/// `(0.1, 10.0)` to `(0.01, 100.0)` so close-but-distinct small
-/// per-segment Sharpes (e.g. 0.05 vs 0.09 → `1/sharpe` of 20 vs 11)
-/// no longer both saturate the upper bound. Under the tight clamp,
-/// every cycle produced near-uniform weights → `curric_conc=0`. The
-/// wider range preserves segment-rank fidelity while still keeping
-/// the post-normalisation weights well-conditioned (max/min ratio
-/// of 1e4 instead of 1e2). Per `pearl_controller_anchors_isv_driven`
-/// these bounds should eventually be ISV-driven too — left as a
-/// follow-up because the immediate Phase 8.2 goal is just to unblock
-/// the dark-code path; full adaptive bounds are Phase 9 scope.
+/// **Phase 8.6 (2026-05-12)** — rewrote from `1/sharpe.max(0.01)` to
+/// `exp(-(sharpe - mean_sharpe) / std_sharpe)`. The old formula
+/// saturated at the upper clamp (100.0) whenever per-segment Sharpe
+/// was near zero or negative — and with `pnl_mean ≈ 1e-7` for a
+/// PF<1 policy, EVERY segment had `|sharpe| < 0.01` → all saturated
+/// to 100 → uniform weights → CV=0. v8 cycle 1 (commit 5694eb4df)
+/// confirmed `curric_conc=0.0000` despite the wider Phase 8.2 clamp
+/// and the more sensitive Phase 8.4 CV formula.
+///
+/// The new formula is z-score-normalised on per-segment Sharpe:
+///
+///   w_i = exp(-(sharpe_i - sharpe_mean) / sharpe_std.max(eps))
+///
+/// Properties:
+///   - **Scale-invariant**: z-score normalisation removes the
+///     per-segment Sharpe magnitude (the killer of the old formula
+///     near zero).
+///   - **Monotonic & smooth**: weight grows as Sharpe drops below
+///     the segment-mean (harder segment → higher curriculum weight),
+///     no saturation cliff.
+///   - **Differential-difficulty preserved at any scale**: weights
+///     spread regardless of whether per-segment Sharpes are
+///     `~1.0` (healthy) or `~1e-3` (degenerate cold-start).
+///   - **Uniform-input handled**: `sharpe_std → 0` falls back to
+///     uniform weights (correct cold-start sentinel).
+///
+/// z-clamp at [-3, 3] before exp keeps weights in `[exp(-3), exp(3)] =
+/// [0.05, 20]` — well-conditioned for normalisation without losing
+/// the differential signal.
+///
+/// Per `pearl_controller_anchors_isv_driven` and
+/// `feedback_isv_for_adaptive_bounds`, the normalisation anchor is
+/// the observed cycle-scoped `sharpe_std` (signal-driven), not a
+/// hardcoded constant.
 fn compute_curriculum_weights(trades: &[EvalTrade], n_segments: usize) -> Vec<f32> {
     if trades.is_empty() || n_segments == 0 {
         return Vec::new();
     }
     let seg_size = (trades.len() / n_segments).max(1);
-    let mut weights: Vec<f32> = (0..n_segments)
+
+    // Pass 1: per-segment Sharpe.
+    let sharpes: Vec<f32> = (0..n_segments)
         .map(|seg| {
             let start = seg * seg_size;
             let end = ((seg + 1) * seg_size).min(trades.len());
             if start >= trades.len() {
-                return 0.01;
+                return 0.0;
             }
             let slice = &trades[start..end];
             let mean: f32 =
@@ -619,8 +644,27 @@ fn compute_curriculum_weights(trades: &[EvalTrade], n_segments: usize) -> Vec<f3
                 .map(|t| (t.pnl - mean).powi(2))
                 .sum::<f32>()
                 / slice.len().max(1) as f32;
-            let sharpe = mean / var.sqrt().max(1e-6);
-            (1.0 / sharpe.max(0.01)).clamp(0.01, 100.0)
+            mean / var.sqrt().max(1e-6)
+        })
+        .collect();
+
+    let n = sharpes.len() as f32;
+    let sharpe_mean: f32 = sharpes.iter().sum::<f32>() / n;
+    let sharpe_var: f32 = sharpes
+        .iter()
+        .map(|s| (s - sharpe_mean).powi(2))
+        .sum::<f32>()
+        / n;
+    let sharpe_std = sharpe_var.sqrt().max(1e-6);
+
+    // Pass 2: z-score normalised exp weights.
+    let mut weights: Vec<f32> = sharpes
+        .iter()
+        .map(|&s| {
+            // Negate so HARDER segments (lower Sharpe) get larger weights.
+            let z = -(s - sharpe_mean) / sharpe_std;
+            // z-clamp to keep exp range well-conditioned.
+            z.clamp(-3.0, 3.0).exp()
         })
         .collect();
     let total: f32 = weights.iter().sum();

docs/dqn-wire-up-audit.md

@@ -15843,3 +15843,94 @@ masking it. A future Phase will either wire PPO/supervised branch
 sizes correctly (PPO uses a 5-exposure action space, supervised uses
 signal thresholds — different conventions from DQN factored actions)
 or delete the dead paths per `feedback_no_partial_refactor`.
+
+## 2026-05-12 — SP21 T2.2 Phase 8.6: curriculum weights → z-score exp(-sharpe) (atomic)
+
+### Scope (atomic single commit)
+
+Rewrote `compute_curriculum_weights` from `1/sharpe.max(0.01)` to
+`exp(-(sharpe - mean)/std)` (z-score normalised). v8 smoke (train-96wfk,
+commit 5694eb4df) cycle 1 confirmed `curric_conc=0.0000` despite the
+Phase 8.2 clamp relaxation and Phase 8.4 CV formula — root cause was
+the per-segment weight function itself, not the post-hoc concentration
+metric.
+
+### Root cause
+
+```rust
+let sharpe = mean / var.sqrt().max(1e-6);
+(1.0 / sharpe.max(0.01)).clamp(0.01, 100.0)   // ← saturates at 100 when sharpe < 0.01
+```
+
+For v8 cycle 1 (PF=0.94, val-trade mean ≈ negative, pnl_std=1.08e-5),
+per-segment Sharpe is well under 0.01 in absolute value. `sharpe.max(0.01) = 0.01`
+→ `1/0.01 = 100` → upper clamp → all 8 segments map to 100 → uniform
+weights after normalisation → CV = 0. The Phase 8.4 CV formula was
+correct but it operates on the post-saturated weights, which destroy
+the differential signal upstream.
+
+### Fix
+
+z-score normalised exp(-sharpe) per segment:
+
+```rust
+let z = -(s - sharpe_mean) / sharpe_std;
+z.clamp(-3.0, 3.0).exp()
+```
+
+Properties:
+- **Scale-invariant**: z-score removes per-segment Sharpe magnitude
+  (the killer near zero).
+- **Monotonic & smooth**: weight grows as Sharpe drops below mean
+  (harder segment → higher curriculum weight); no saturation cliff.
+- **Differential-difficulty preserved at any scale**: weights spread
+  whether per-segment Sharpes are `~1.0` (healthy) or `~1e-3`
+  (degenerate cold-start).
+- **Uniform-input handled**: `sharpe_std → 0` falls back to uniform
+  weights via the `.max(1e-6)` floor + post-norm (correct sentinel).
+- **z-clamp [-3, 3]** keeps `exp` range `[0.05, 20]` — well-
+  conditioned without losing differential signal.
+
+### Why win_conc=0 is NOT addressed here
+
+Phase 8.4 noted that `win_conc=0` for v8 cycle 1 was the
+`compute_winner_concentration` guard intentionally tripping when
+`all_mean ≤ 0.01×pnl_std`. With PF=0.94 the val-trade mean is genuinely
+negative → guard is correctly suppressing "concentration of wins" for
+a losing policy. This is honest reporting, not a bug. A future Phase
+may revisit whether to emit a non-zero signal for losing policies
+(e.g., negative concentration as "anti-signal"), but that's an
+architectural decision, not a scale/saturation fix.
+
+### Files changed
+
+- `crates/ml/src/trainers/dqn/trainer/enrichment.rs`:
+  `compute_curriculum_weights` rewritten (z-score exp); two-pass
+  per-segment Sharpe → z-score → exp(-z) → normalise.
+- `docs/dqn-wire-up-audit.md`: this entry
+
+### Pearls + invariants honoured
+
+- **pearl_controller_anchors_isv_driven**: normalisation anchor is
+  the observed `sharpe_std` (signal-driven), not magic constants
+- **feedback_isv_for_adaptive_bounds**: z-clamp `[-3, 3]` is structural
+  (3σ tail-cutoff for numerical safety), not a tuned value
+- **pearl_first_observation_bootstrap**: uniform Sharpe → std=0 →
+  z=0 → uniform weights (correct cold-start sentinel)
+- **feedback_no_quickfixes**: structural reformulation (1/x → exp(-z))
+  rather than a clamp adjustment
+
+### Verification
+
+```
+cargo check -p ml --features cuda           # clean
+```
+
+### Expected v9 smoke
+
+For v9 cycle 1 (mirror v8's per-segment Sharpes in [-0.01, 0.01]
+range with std ~0.005):
+- z-scores roughly `[-2, 2]` across segments
+- weights roughly `[0.14, 7.4]` pre-normalisation
+- post-normalisation: max/min ratio ~50× — strong differential signal
+- `curric_conc` (Phase 8.4 CV formula): should reach ~0.3-0.6